For production environments, a hybrid architecture of edge preprocessing plus cloud inference is recommended, with the edge handling video frame extraction, ROI cropping, and preliminary filtering, and the cloud running heavier detection models and aggregating multi-camera data.